A computing apparatus to classify anomalies in images, by unsupervised anomaly detection on an input dataset of the images to detect anomaly portions from said images to generate, for an image in the dataset, a corresponding mask image transmitting a detected anomaly portion in the image and blocking anomaly-free portions; train a classifier ANN, including, in a first epoch process processing a masked version of the input dataset with the classifier ANN, the masked version including the image of the input dataset masked by the corresponding mask image, and training the classifier ANN to classify anomaly portions into one of plural classes by minimising a cross entropy loss function using generated labels as ground truths; extracting, from the classifier ANN, a latent feature representation of the image of the masked version in the input dataset; and in a second epoch process generating a set of pseudo labels corresponding to the masked version of the input dataset by applying an unsupervised clustering algorithm to the latent feature representations to cluster the latent feature representations into one of plural clusters each with a different associated pseudo label, to obtain a pseudo label corresponding to the image in the input dataset; training the classifier ANN to minimise a loss function between a class into which the image of the input dataset is classified by the classifier ANN using the pseudo label for said image as ground truth The training epoch may be repeatedly executed until satisfaction of a training condition, and to output, for the image in the input dataset, an identification of the detected anomaly portion with a corresponding class into which the anomaly portion is classified by the classifier ANN.
Legal claims defining the scope of protection, as filed with the USPTO.
2. The computing apparatus according to claim 1, wherein the masked version of the image in the input dataset processed by the classifier ANN in the first process, in addition to being masked by the corresponding mask image, is filtered by an image filter or transformed by an image transform algorithm.
3. The computing apparatus according to claim 2, wherein the image filter is an emboss image filter or the image transform algorithm is an embossing image transform algorithm.
4. The computing apparatus according to claim 2, wherein the masked version of the image in the input dataset comprises plural versions of the image in the input dataset, the plural versions being the image filtered with a selection of a filter among filters including plural image transform algorithm or image filters applied.
5. The computing apparatus according to claim 4, wherein the automatically generated labels are determined by the image transform algorithm or an image filter applied to a version, among the plural versions, so that the image transform algorithm or the image filter maps to a label among label values, the label values being arbitrary values different from one another.
6. The computing apparatus according to claim 1, wherein the unsupervised anomaly detection is performed with an autoencoder or a generator neural network, pre-trained to generate defect-free portions of images, and to generate a mask image corresponding to an image in the input dataset by generating a defect-free version of the image, and comparing the image with the defect-free version of the image to obtain the mask image.
7. The computing apparatus according to claim 6, wherein the generator neural network is a generative adversarial neural network.
8. The computing apparatus according to claim 1, wherein the loss function in the first process is a pixel-wise cross entropy loss.
10. A system, comprising the computing apparatus according to claim 1, and imaging apparatus configured to generate the images and to store the images as the input dataset for processing by the computing apparatus.
11. The system according to claim 10, wherein the images are images of production samples, and further comprising a production environment for producing products or materials as production samples.
13. The method according to claim 12, wherein the masked version of the image in the input dataset processed by the classifier ANN in the first process, in addition to being masked by the corresponding mask image, is filtered by an image filter or transformed by an image transform algorithm.
14. The method according to claim 13, wherein the image filter is an emboss image filter or the image transform algorithm is an embossing image transform algorithm.
15. The method according to claim 13, wherein the masked version of the image in the input dataset comprises plural versions of the image in the input dataset, the plural versions being the image filtered with a selection of a filter among filters including plural image transform algorithm or image filters applied.
16. The method according to claim 15, wherein the automatically generated labels are determined by the image transform algorithm or an image filter applied to a version, among the plural versions, so that the image transform algorithm or the image filter maps to a label among label values, the label values being arbitrary values different from one another.
17. The method according to claim 12, wherein the unsupervised anomaly detection is performed with an autoencoder or a generator neural network, pre-trained to generate defect-free portions of images, and to generate a mask image corresponding to an image in the input dataset by generating a defect-free version of the image, and comparing the image with the defect-free version of the image to obtain the mask image.
18. The method according to claim 17, wherein the generator neural network is a generative adversarial neural network.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
December 10, 2021
September 24, 2024
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.